fix(workflows): reject a non-string prompt in prompt-step validate() (#3582)

`PromptStep.execute` str()-coerces `config['prompt']` and dispatches the
result to the integration CLI as the model's instructions. But its `validate`
only checked that `prompt` was *present*, not that it was a string — the exact
parity gap the sibling `ShellStep` closes for `run`.

So a YAML authoring slip like `prompt: [review, this]` or `prompt:` (null)
passed validation, then `execute` sent the Python repr (`"['review', 'this']"`,
`"None"`) to the LLM verbatim — silently wrong instructions with no error and a
COMPLETED status. The engine does not auto-validate step config
(`load_workflow` explicitly defers validation), so validation is the only place
this surfaces before dispatch.

Extend `validate` to reject any non-string `prompt` with the shell-step's
phrasing ("'prompt' must be a string, got <type>"), mirroring the shell `run`
and command `input`/`options` type checks. A `{{ ... }}` expression is still a
str, so it stays valid. Adds regression coverage for non-string prompts
(null/list/int/dict) and confirms an expression prompt still validates.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

src/specify_cli/workflows/steps/prompt/__init__.py

@@ -160,4 +160,16 @@ class PromptStep(StepBase):
             errors.append(
                 f"Prompt step {config.get('id', '?')!r} is missing 'prompt' field."
             )
+        elif not isinstance(config["prompt"], str):
+            # execute() str()-coerces prompt and dispatches it to the
+            # integration CLI, so a null or list 'prompt' would send the Python
+            # repr ('None', "['review', 'this']") to the model as instructions —
+            # silently wrong, with no error. Reject non-strings at validation,
+            # mirroring the shell-step 'run' and command-step input/options type
+            # checks. An expression like "{{ ... }}" is still a str, so it stays
+            # valid.
+            errors.append(
+                f"Prompt step {config.get('id', '?')!r}: 'prompt' must be a "
+                f"string, got {type(config['prompt']).__name__}."
+            )
         return errors

tests/test_workflows.py

@@ -1381,6 +1381,21 @@ class TestPromptStep:
         errors = step.validate({"id": "test"})
         assert any("missing 'prompt'" in e for e in errors)
 
+    @pytest.mark.parametrize("bad_prompt", [None, ["review", "this"], 42, {"a": 1}])
+    def test_validate_rejects_non_string_prompt(self, bad_prompt):
+        """A non-string 'prompt' must be rejected at validation.
+
+        execute() str()-coerces prompt and dispatches it to the integration
+        CLI, so a null or list prompt would otherwise send the Python repr to
+        the model as instructions — silently wrong. Mirrors the shell-step
+        'run' type check.
+        """
+        from specify_cli.workflows.steps.prompt import PromptStep
+
+        step = PromptStep()
+        errors = step.validate({"id": "p", "prompt": bad_prompt})
+        assert any("'prompt' must be a string" in e for e in errors)
+
     def test_validate_valid(self):
         from specify_cli.workflows.steps.prompt import PromptStep
 
@@ -1388,6 +1403,16 @@ class TestPromptStep:
         errors = step.validate({"id": "test", "prompt": "do something"})
         assert errors == []
 
+    def test_validate_accepts_expression_prompt(self):
+        """A '{{ ... }}' expression prompt is a str, so it stays valid."""
+        from specify_cli.workflows.steps.prompt import PromptStep
+
+        step = PromptStep()
+        errors = step.validate(
+            {"id": "p", "prompt": "Review {{ inputs.file }}"}
+        )
+        assert errors == []
+
 
 class TestShellStep:
     """Test the shell step type."""